The traditional hyperspectral unmixing methods only consider the geological properties of hyperspectral images or the sparse properties of abundance and neglect the spectral spatial information of hyperspectral data. Thus when the pure pixels are missing, the unmixing accuracy is significantly reduced. In order to overcome these limitations, an improved spatial information constrained nonnegative matrix factorization method for unmixing is proposed. This method fully uses the spatial information and the sparse properties of hyperspectral images, and thus the properties of the traditional nonnegative matrix factorization methods are improved. Both the synthetic simulation images and the experimental results show that the proposed method has overcome the noise-sensitivity and the dependence on pure pixels of the traditional methods.